Structure to Energy and Forces on OC20 S2EF (test splits average)
19Force MAE (meV/Å)GemNet-OC-L-F
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GemNet-OC-L-FTrain set=OC20 + OC-MD, Throughput (Samples / GPU sec)=3.22022.04 | 19 | 241 | 0.691 | 2.97 | |
| GemNet-XLTrain set=OC20, Throughput (Samples / GPU sec)=1.52022.04 | 20.5 | 270 | 0.66 | 1.81 | |
| GemNet-OCTrain set=OC20, Throughput (Samples / GPU sec)=18.32022.04 | 20.7 | 233 | 0.666 | 2.5 | |
| GemNet-OC-L-ETrain set=OC20 + OC-MD, Throughput (Samples / GPU sec)=7.52022.04 | 21 | 230 | 0.665 | 2.8 | |
| GemNet-dTTrain set=OC20, Throughput (Samples / GPU sec)=25.82022.04 | 24.2 | 292 | 0.616 | 1.2 | |
| GemNet-OCTrain set=OC-2M2022.04 | 24.3 | 274 | 0.603 | 1.25 | |
| GemNet-dTTrain set=OC-2M2022.04 | 28.1 | 323 | 0.559 | 0.69 | |
| SpinConvHidden dim=256, #Msg layers=3, #Params=8.5M, Train time=275d, Inference time=9.1h, mode=force-centric2021.06 | 29.66 | 336.3 | 0.5391 | 45 | |
| SpinConvTrain set=OC20, Throughput (Samples / GPU sec)=62022.04 | 29.7 | 336 | 0.539 | 0.45 | |
| ForceNet-largeHidden dim=768, #Msg layers=7, #Params=34.8M, Train time=194d, Inference time=3.5h2021.06 | 31.13 | — | 0.5195 | — | |
| DimeNet++-L-F+ETrain set=OC20, Throughput (Samples / GPU sec)=4.62022.04 | 31.3 | 480 | 0.544 | 0 | |
| ForceNet-largeTrain set=OC20, Throughput (Samples / GPU sec)=15.32022.04 | 32 | — | 0.516 | 0.01 | |
| SpinConvHidden dim=256, #Msg layers=3, #Params=6.1M, Train time=380d, Inference time=22.7h, mode=energy-centric force-only2021.06 | 32.58 | — | 0.4976 | — | |
| DimeNet++-largeHidden dim=512, #Msg layers=3, #Params=10.7M, Train time=1600d, Inference time=27.0h2021.06 | 32.75 | — | 0.5408 | — | |
| PaiNNTrain set=OC20, Throughput (Samples / GPU sec)=602022.04 | 33.1 | 341 | 0.491 | 0.46 | |
| ForceNetHidden dim=512, #Msg layers=5, #Params=11.3M, Train time=31d, Inference time=1.3h2021.06 | 34.32 | — | 0.477 | — | |
| SpinConvTrain set=OC-2M2022.04 | 35.5 | 401 | 0.475 | 0.13 | |
| DimeNet++Hidden dim=192, #Msg layers=3, #Params=1.8M, Train time=587d, Inference time=8.5h, mode=force-only2021.06 | 35.73 | — | 0.4785 | — | |
| SpinConvHidden dim=256, #Msg layers=3, #Params=6.1M, Train time=275d, Inference time=22.7h, mode=energy-centric2021.06 | 38.88 | 411.4 | 0.4299 | 16 | |
| DimeNet++Hidden dim=192, #Msg layers=3, #Params=1.8M, Train time=587d, Inference time=8.5h2021.06 | 47.58 | 534.3 | 0.356 | 5 | |
| SchNetHidden dim=1024, #Msg layers=5, #Params=9.1M, Train time=194d, Inference time=0.8h2021.06 | 49.03 | — | 0.3413 | 0 | |
| SchNetTrain set=OC20, Throughput (Samples / GPU sec)=5492022.04 | 54.7 | 540 | 0.302 | 0.06 | |
| DimeNet++Train set=OC-2M2022.04 | 63 | 761 | 0.222 | 0.01 | |
| CGCNNTrain set=OC202022.04 | 73.3 | 608 | 0.146 | 0.01 | |
| SchNetTrain set=OC-2M2022.04 | 77.1 | 1,370 | 0.116 | 0 | |
| Median2021.06 | 84.38 | 2,258 | 0.0156 | 0.5 | |
| DimeNet++Hidden dim=192, #Msg layers=3, #Params=1.8M, Train time=587d, Inference time=8.5h, mode=energy-only2021.06 | 345.9 | 480.2 | 0.1021 | 0 |